Source Free Domain Adaptation with Pseudo-Labeling Quality Assessed by Sam in Fundus Image Segmentation

Precise segmentation of the optic cup and optic disc in fundus images is crucial for measuring the cup-to-disc ratio, a key indicator for glaucoma screening and diagnosis. When applying a network trained on data from one medical site to data from another site, performance declines due to domain shift problems. This study centers on the realistic scenario of source-free unsupervised domain adaptation. We propose a new pseudo-labeling strategy that takes advantage of the capabilities of the SAM segmentation foundation model by using it to assess the quality of the pseudo-label segmentation of the target images created by the source model. We implemented our method on publicly available retinal segmentation datasets collected from several medical sites and reported improved performance compared to previous methods.